I'm building a visual search system for fashion that automatically detects clothing items in any image and finds similar products in your catalog.
*Technical approach:*
- 8-service microservices architecture with Docker Compose
- GPU-accelerated ML service using NVIDIA Docker runtime
- YOLOv8 for object detection, CLIP for embeddings + structured labeling
- Multi-storage strategy: PostgreSQL (metadata), ChromaDB (vectors), MinIO S3 (images)
- Async processing with Celery workers + Redis broker + Flower monitoring
- Traefik reverse proxy with automatic service discovery + health checks
*Key insight:* Most fashion visual search requires manual product photography. This system works with any image - street style, social media posts, etc. Upload a photo of someone wearing clothes, and it automatically crops and indexes each item.
*Development focus:*
- Separated ML inference into stateless service for better scaling
- Used async job queues to keep API responsive during processing
- Vector embeddings stored in ChromaDB for fast similarity search
- Everything containerized and *runs completely locally* - no external APIs
*Current state:* Core functionality works, but still optimizing crop quality and fine-tuning the ML pipeline. Anyone can clone and run `docker-compose up` to try it.
*Interesting challenges I'm working on:*
- Handling variable image quality and lighting conditions
- Balancing detection accuracy vs processing speed
- Designing async workflows for multi-step ML pipelines
- Service orchestration and dependency management
Would love feedback from the community, especially on approaches to crop quality filtering or experiences with CLIP fine-tuning for domain-specific applications.
*Technical approach:* - 8-service microservices architecture with Docker Compose - GPU-accelerated ML service using NVIDIA Docker runtime - YOLOv8 for object detection, CLIP for embeddings + structured labeling - Multi-storage strategy: PostgreSQL (metadata), ChromaDB (vectors), MinIO S3 (images) - Async processing with Celery workers + Redis broker + Flower monitoring - Traefik reverse proxy with automatic service discovery + health checks
*Key insight:* Most fashion visual search requires manual product photography. This system works with any image - street style, social media posts, etc. Upload a photo of someone wearing clothes, and it automatically crops and indexes each item.
*Development focus:* - Separated ML inference into stateless service for better scaling - Used async job queues to keep API responsive during processing - Vector embeddings stored in ChromaDB for fast similarity search - Everything containerized and *runs completely locally* - no external APIs
*Current state:* Core functionality works, but still optimizing crop quality and fine-tuning the ML pipeline. Anyone can clone and run `docker-compose up` to try it.
*Interesting challenges I'm working on:* - Handling variable image quality and lighting conditions - Balancing detection accuracy vs processing speed - Designing async workflows for multi-step ML pipelines - Service orchestration and dependency management
Would love feedback from the community, especially on approaches to crop quality filtering or experiences with CLIP fine-tuning for domain-specific applications.